New models for sustained work, coding agents you can try, and the cloud systems that run them. This first edition establishes a baseline; later reports will focus on meaningful changes since the previous edition.
Edition 01 · Coverage: 7-14 October 2026
Details on the two latest updates. Open a radar item for its announcement.
A model built for long coding tasks, including migrations across a codebase.
Google opened developer access on the paid Gemini API and in Antigravity for Google AI Ultra subscribers.
Try it on a multi-file migration with tests that can verify the result. The relevant question is whether it completes your repository task reliably.
Access Paid Gemini API; Antigravity for Google AI Ultra subscribers.
Models for coding and personal tasks that recall context and work through blockers.
Meta launched two related models emphasizing retrieval of relevant memories and sustained execution across difficult tasks. The family name groups Wonder and Wonder Full.
Evaluate continuity across sessions and recovery from a failed approach, alongside code correctness.
Access Wonder: Meta Model API and Muse Code. Wonder Full: Muse Code subscribers.
Open-source projects with a concrete reason to pay attention.
The license badges apply to the linked software. Model providers, commercial agents and hosted services may have separate access requirements and costs.
Write repository tasks in Markdown and run an AI agent through GitHub Actions.
This pre-release tightens MCP controls, rejects unsafe pull_request_target roles in strict mode and changes model-routing rules.
Existing workflows may need changes before recompilation; review the release before upgrading.
GitHub CLI, a repo you can edit, Actions enabled and supported model authentication.
Agent Canvas coordinates coding agents and automations across execution backends.
v1.27.0 adds OpenCode and Pi as local ACP agents; v1.27.1 fixes Docker bridge and theme issues.
A useful watch if you want one control center for several coding agents.
Self-host locally or with Docker; bring the model or agent access you want to use.
An open-source terminal coding agent with a beta desktop app and configurable model providers.
Adds explicit upgrade paths toward OpenCode 2, an /update command and tighter terminal-session checks. This version is not the OpenCode 2 release.
Track it for provider choice, permission changes and how the major-version migration develops.
Install the CLI or desktop app, then connect your model provider.
A stateful agent harness that can delegate coding work to an OpenCode worker.
Adds explicit OpenCode workers and a cloud-only webhook-management skill. The repository already existed; the worker integration is the fresh capability.
Try persistent coordination between harnesses. Workers must be invoked explicitly and OpenCode must be installed on PATH.
Local mode is available with your provider. Webhook management requires Letta Cloud.
Context, execution and the infrastructure around agents.
An update to the existing Notion product, with shared agent context, tighter access controls and richer edit history.
The release includes per-person MCP access revocation, database column permissions and agent-attributed version history, alongside expanded connections.
Keep shared context in one place, restrict sensitive fields and review agent edits. Column permissions require Business or Enterprise.
The Skills API was announced on 17 September; its appearance in this release is a recap, not a new API launch.
Google's new Gemini agent announcement overlaps with Claude Managed Agents, which has run custom cloud agents since April.
Google announced its enterprise agent on 8 October. Anthropic added dynamic workflows on 9 October: an agent writes a program to coordinate agents in phases, and the server runs it in the background.
Choose between an enterprise work assistant and a developer service for custom agents; persistence alone is already offered by both.
Google's whole-product rollout status was not established by the retrieved primary announcement. Claude Managed Agents is public beta.
Recent papers and publications, with their findings and limits.
Same Attack, Different Judge: A First Controlled Adversarial Comparison of Jev, LLM Judges, and Purpose-Built Safety Classifiers
Tests how Jev, LLM judges and safety classifiers handle the same adversarial inputs across moderation, routing and phishing tasks.
The authors find that robustness varies by task: Jev resists structured forgeries in moderation but is more vulnerable in phishing.
Test an agent’s routing or safety judge on your own workload, including benign inputs and false alarms.
One Jev version and one English phishing corpus; the study does not establish a universal winner.
AgentSWE: Can Coding Agents Build the Agent You Actually Want?
Benchmarks whether coding agents can create, edit and improve other agents that meet real workflow requirements.
The 25-task benchmark reports overfitting to visible examples, incomplete runtime behavior and unsupported claims of completion.
Freeze the delivered agent and test unseen acceptance cases, including its runtime behavior.
Main builder/task pairs ran once; some scores rely on model judges.
Useful comparisons for deciding what to try.
They overlap in hosted, long-running work. The product-scope distinction below is an editorial synthesis of their documented surfaces.
| Dimension | Google Gemini agent | Claude Managed Agents |
|---|---|---|
| Best fit | Enterprise work assistant across workplace apps, with app and API access. | Developer service for building custom Claude-powered agents. |
| Models | Gemini and Claude orchestration; more model support is planned. | Claude 4.5 and later; different supported Claude models per agent. |
| Long-running work | Cloud work, shared context and dynamically created helper agents. | Persistent sessions and filesystem state; orchestration and shared memory. |
| Controls | Workplace identity, authorization, gateway policy and audit logs. | Tool permissions, credential vaults, network restrictions, budgets and events. |
| Availability | Announced 8 Oct; broad availability unverified in the primary announcement. | Public beta; some capabilities remain restricted research previews. |
Three implementations of the same broad idea: return a bounded decision that software can use, such as a choice, a score or a probability.
Dedicated decision model with its own architecture, parallel sampler and training for calibrated decisions.
Primary source ↗Separate Microsoft model post-trained from Qwen3.5-9B to score predefined choices; Foundry and OpenRouter access.
Primary source ↗Dedicated endpoint backed by GPT-6 Luna, accepting text and images and returning typed decisions; public beta.
Primary source ↗One practical experiment from today’s briefing.
GitHub Agentic Workflows runs repository tasks from Markdown instructions through GitHub Actions. Give it a crash report, confusing setup instructions and a feature request; it should produce one linked triage report.
Why it helps See whether an agent can save time sorting incoming issues and deciding what to investigate first. Compare its classifications, priorities and follow-up questions with your own judgement.
Open the step-by-step guideCreate three sample issues in a disposable personal repository.
Add the provided gh-aw Markdown workflow and compile it to Actions.
Run it manually, inspect the report and score the three classifications.
GitHub CLI, Actions and Copilot authentication. First-time setup may take longer than 30 minutes.
The custom workflow compiled successfully with official v0.89.21 and v0.91.12 binaries. An actual GitHub Actions/model run has not been performed.